A semantic computer interface

Semwright gives agents a structured way to work with a computer. It exposes typed operations through a common broker, rather than making screenshots or unrestricted shell access the default interface.

The project is a development preview, starting with Linux. It is not a model, a hosted agent, or a production-certified desktop automation service.

Start in a disposable environment

Build from source, run the fake-backend smoke path, and then inspect the first commands. Do not attach an unreviewed build to a credential-rich desktop.

Understand the path

An agent discovers an operation and supplies structured inputs. The broker checks capability, policy, references and execution context before routing the request. Results carry execution provenance.

CLI / MCP / recipes / inspector
               ↓
      broker · policy · audit
               ↓
        provider runtime
               ↓
 native APIs / app drivers / MCP providers

Changing the frontend does not create a more privileged route. Discovery describes what a capability can do; it does not authorize its use.

Choose a reading path

Agent developers can connect an MCP client and learn capabilities and references. Systems engineers can inspect the architecture and permissions. Application developers can explore drivers and providers.

The public website is static. Its examples never connect to a local broker or execute commands on your computer.

Source: architecture at the reviewed revision. Reviewed 22 September 2026.